A group of startups working in technically deep fields stood out at Y Combinator’s latest Demo Day, ranging from floating data centers and nuclear energy to artificial intelligence chips, robotics and application programming interfaces for controlling robots. The list was based on recommendations from investors whom TechCrunch asked to identify the most interesting companies in the summer batch, including companies mentioned by at least two investors.
According to the report, this batch appeared more inclined toward deep technologies than previous batches, while one investor described its ideas as closer to science fiction. At the same time, investors considered the financial valuations more realistic compared with those of other recent batches.
Artificial Intelligence Infrastructure and Energy
Automarine, also referred to elsewhere in the text as Atomarine, aims to build nuclear-powered data centers on floating platforms at sea. The company is betting on seawater cooling, away from power constraints and local communities’ objections to building new data centers. It plans to launch a gas-powered prototype in 2028, then transition to nuclear-powered floating vessels in 2032, and says it has secured letters of intent representing customer interest worth more than $4 billion.
Dipole Labs is developing a photonic switch for artificial intelligence data centers, with the goal of keeping data in optical form as it moves between chips instead of converting it back and forth between light and electrical signals. The company believes this could reduce energy consumption and heat while improving the utilization of graphics processing unit clusters.
Lamb Labs is working on custom inference chips that embed artificial intelligence model weights directly into silicon. The company calls these chips “model processing units” (MPUs) and says they avoid the bottleneck caused by transferring weights from memory while models are running.
Robots for Industrial and Household Use
Praxis AI collects videos and data about people’s performance of real-world tasks, then converts them into training materials for companies developing robots. It says it is already working with publicly traded companies and has collected data from more than 150 different environments.
Nori offers a household robot priced at approximately $1,600 that can be operated through an application on a laptop and is designed to clean and fold clothes. The company, which launched six weeks before the report was prepared, says it has generated nearly half a million dollars in sales. The report compares this price with the approximately $20,000 price of a Neo robot, while the central question remains whether these devices can actually perform complex household tasks.
Cosmic Robotics is developing autonomous robots capable of lifting heavy loads. It says its technology is being used to install solar panels in the United States and that it has contracts worth $25 million through 2027. Its founders connect the project to a vision of building a city on Mars, targeting the start of an exploratory mission in 2028.
From Drones to Biological Computing
Isengard Industries wants to produce offensive and counter-drone aircraft powered by jet engines within allied countries, at a cost lower than the prices charged by major U.S. defense contractors. The company says it generates $10 million in revenue and received a high valuation within the batch, according to two investors.
Parasma is exploring the use of human brain cells to power computing with greater energy efficiency, while Waddle Labs is developing an API layer that uses large language model agents to write robot-control code directly. The company says developers will be able to issue commands in natural language, while the agents create the executable code and verify that it works, with the robot set up in approximately 20 minutes.
Why Does This Trend Matter?
The list reveals a shift in the attention of some early-stage capital away from artificial intelligence applications alone and toward the layers that make them possible: energy, networks, chips, data and robots. However, most of the figures cited, such as letters of intent, contracts, sales and valuations, are based on statements from the companies or descriptions by investors, and do not necessarily represent realized revenue or proven technical success. Scalability, reliability, safety and regulatory compliance therefore remain open questions for these projects.